iSAID数据集

[SherryZeak]

Hello,

Based on your description, I attempted to preprocess the iSAID dataset using "mmsegmentation". However, the structure of the processed dataset is as follows:
iSAID
├── img_dir
│ ├── train
│ ├── val
│ ├── test
├── ann_dir
│ ├── train
│ ├── val

In your code ConvLSR-Net/config/isaid/convlsrnet.py, the paths to the dataset are specified as follows:

train_dataset = ISAIDDataset(
data_root="/home/zrh/datasets/semantic/isaid/train",
mode='train',
mosaic_ratio=0.4,
transform=train_aug
)
val_dataset = ISAIDDataset(
data_root="/home/zrh/datasets/semantic/isaid/val",
transform=val_aug
)
test_dataset = ISAIDDataset(
data_root="/home/zrh/datasets/semantic/isaid/val",
transform=val_aug
)
I have noticed some discrepancies between these structures, which are causing errors during processing. I have tried various methods to adjust the setup, but the issues persist. Could you please provide the structure of your processed dataset located at /home/zrh/datasets/semantic/isaid or share the related preprocessing steps you used?

Disclaimer: This request is solely for academic research and comparison purposes and will not be used for commercial applications.

Thank you for your assistance!

Best regards,

[stdcoutzrh]

Hello, first of all, sorry for the missing description in the dataset processing section that hindered your experiment.
According to the processing results in mmsegmentation, iSAID will be obtained:
├── img_dir
│ ├── train
│ ├── val
│ ├── test
├── ann_dir
│ ├── train
│ ├── val
In order to reuse the processing code of the dataset in Geoseg, it is necessary to create new 'train' and 'val' directories.
Change the name of 'train' in 'img_dir' to 'images', and change the name of the 'train' in 'ann_dir' to 'masks', and move these two in the new 'train' directory.
Change the name of 'val' in 'img_dir' to 'images', and change the name of the 'val' in 'ann_dir' to 'masks', and move these two in the new 'val' directory.
If you have any further questions, please continue to leave a message. Thank you.

[SherryZeak]

Thank you for your generous explanation. Our team has successfully replicated the experiments related to Vaihingen and Potsdam, and the final results are generally consistent with yours.

Based on your iSAID data, we have:

• Renamed the 'train' folder in 'img_dir' to 'images' and renamed the 'train' folder in 'ann_dir' to 'masks', then moved these two directories into a new 'train' directory.
• Renamed the 'val' folder in 'img_dir' to 'images' and renamed the 'val' folder in 'ann_dir' to 'masks', then moved these two directories into a new 'val' directory.

However, when I execute the training command python train_supervision.py -c ./config/isaid/convlsrnet.py, an error occurs. The main error message is:

FileNotFoundError: [Errno 2] No such file or directory: '/home/zeak/anaconda3/envs/air/ao/data/iSAID/val/images/P2598_1024_1920_0_896_instance_color_RGB.png'

Our team has strictly followed your instructions to create the new 'train' and 'val' directories. Additionally, we have updated the dataset paths in the ConvLSR-Net/geoseg/datasets/isaid_dataset.py and ConvLSR-Net/config/isaid/convlsrnet.py files accordingly.